[Paper Review] Deep Dense and Convolutional Autoencoders for Unsupervised Anomaly Detection in Machine Condition Sounds

Alexandrine Ribeiro, Luís Miguel Matos+5 more|arXiv (Cornell University)|Jun 18, 2020

Music and Audio ProcessingComputer Science14 references21 citations

TL;DR

This paper proposes deep dense and convolutional autoencoders (AEs) for unsupervised anomaly detection in machine condition sounds using mel-spectrogram features. Trained only on normal sound samples, the models detect anomalies by measuring reconstruction error, achieving competitive performance—particularly the CNN AE on slider and valve machines—outperforming the DCASE 2020 baseline across all metrics.

ABSTRACT

This technical report describes two methods that were developed for Task 2 of the DCASE 2020 challenge. The challenge involves an unsupervised learning to detect anomalous sounds, thus only normal machine working condition samples are available during the training process. The two methods involve deep autoencoders, based on dense and convolutional architectures that use melspectogram processed sound features. Experiments were held, using the six machine type datasets of the challenge. Overall, competitive results were achieved by the proposed dense and convolutional AE, outperforming the baseline challenge method.

Motivation & Objective

• Address unsupervised anomaly detection in machine condition sounds using only normal training data.
• Develop deep autoencoder models that generalize across diverse machine types without prior anomaly knowledge.
• Improve upon the DCASE 2020 challenge baseline by leveraging deeper architectures and mel-spectrogram features.
• Evaluate model performance across six distinct machine types (ToyCar, ToyConveyor, fan, pump, slider, valve) using standard metrics.
• Design a hybrid approach selecting the best AE per machine type to maximize overall detection performance.

Proposed method

• Use mel-spectrogram features as input to both dense and convolutional autoencoders for temporal-spectral representation of sound.
• Implement a deep fully-connected autoencoder with four 512-unit hidden layers, batch normalization, ReLU activation, and an 8-unit bottleneck layer.
• Design a convolutional autoencoder with strided convolutions and transposed convolutions for hierarchical feature learning in spectrograms.
• Train both models exclusively on normal sound samples to minimize reconstruction error on normal patterns.
• Detect anomalies by computing reconstruction error on unseen data—higher error indicates anomaly.
• Apply a mixed model strategy that selects the best-performing AE (dense or CNN) per machine type based on validation performance.

Experimental results

Research questions

• RQ1 Can deep dense and convolutional autoencoders effectively detect anomalous machine sounds using only normal training data?
• RQ2 How do mel-spectrogram-based autoencoders compare to the DCASE 2020 baseline in terms of AUC and pAUC across diverse machine types?
• RQ3 Does the convolutional autoencoder outperform the dense autoencoder in capturing spatial-temporal patterns in sound spectrograms?
• RQ4 Can a hybrid model combining the best AE per machine type improve overall anomaly detection performance?
• RQ5 What is the impact of architectural depth and normalization (batch norm) on reconstruction accuracy and anomaly detection robustness?

Key findings

• The dense autoencoder outperformed the DCASE 2020 baseline on all six machine types in terms of mean AUC (78.77% vs. baseline) and pAUC (67.58% vs. baseline).
• The convolutional autoencoder achieved the highest mean AUC (80.79%) and pAUC (71.17%) across all models, outperforming the baseline on five of six machine types.
• For the slider machine type, the convolutional autoencoder achieved an AUC of 98.86% and pAUC of 94.47%, indicating excellent anomaly discrimination.
• The mixed model approach, selecting the best AE per machine type, yielded the strongest overall performance, with mean AUC of 83.45% and pAUC of 73.50%.
• The dense autoencoder achieved the best results on ToyCar (AUC 85.97%), ToyConveyor (AUC 76.43%), fan (AUC 72.03%), and pump (AUC 73.06%), while the CNN AE excelled on slider (AUC 91.77%) and valve (AUC 78.83%).
• The best-performing model (CNN AE on slider) achieved an AUC of 98.86%, indicating near-perfect anomaly detection capability on that machine type.